Policy works by governing governance: regulators should treat AI interventions as instruments that reshape decision processes and institutional coordination, not as one‑off fixes to social harms; practical design should focus on incentives, information flows, and adaptive institutions.

Main Finding

Public policy should be understood not as acting directly on “social problems” but as a form of second‑order governance that intervenes in the social management system — the emergent, interdependent pattern of management decisions societies use to respond to problematized situations. Policy problems are decision‑oriented constructions that render complex situations governable; policy therefore reorients how actors make decisions rather than definitively “solving” substantive social issues.

Key Points

• Problem instability: Policy problems are unstable, contested, and resistant to definitive solutions; this is not only due to technical or political obstacles but to a deeper ontological confusion about what policy can target.
• Decision‑oriented constructions: “Problems” are constructed to enable decision making; they are tools for governance rather than pre‑existing objects that policy can straightforwardly fix.
• Social management system: Instead of acting on social systems directly, policy acts on the emergent pattern of management decisions (the social management system) that collectively shape societal responses.
• Second‑order governance: Policy is reconceived as governance of governance — it shapes incentives, norms, information, and decision architectures that steer other actors and institutions.
• Implication for effectiveness: Expect policy to reorient behavior and institutional responses rather than produce final technical solutions to complex social issues.

Data & Methods

• The article is theoretical and conceptual rather than empirical.
• Main method: synthesis and theoretical argumentation using social systems theory as the frame of reference; engages existing literatures on problem definition, framing, wicked problems, and governance.
• No original quantitative data or empirical testing is reported; claims are normative/analytic and meant to reframe how policy interventions are conceptualized.

Implications for AI Economics

• Reframe AI policy as second‑order interventions:
  • AI regulation and public interventions should be designed to reorient firms’, researchers’, regulators’, and users’ decision processes (incentives, disclosure practices, standards, monitoring), not assumed to directly “solve” harms.
• Modeling and empirical work:
  • Economic models of AI policy should treat policy variables as affecting decision architectures and coordination among heterogeneous agents (firms, platforms, institutions), not only as exogenous shocks to production or costs.
  • Use methods suited to emergent, interdependent dynamics: agent‑based models, system dynamics, network analysis, and institutional econometrics to capture how policies reshape management patterns over time.
• Measurement and evaluation:
  • Evaluate policy by tracking changes in decision processes and governance patterns (e.g., corporate governance of AI, procurement rules, standard adoption, information flows) rather than single outcome targets alone.
  • Design metrics for second‑order outcomes: adoption of safety practices, changes in internal incentive structures, interoperability of oversight mechanisms, and coordination among regulators.
• Policy design:
  • Favor adaptive, iterative, and polycentric governance instruments (regulatory sandboxes, standards bodies, conditional approvals, monitoring and reporting requirements) that steer behaviors and allow recalibration as management patterns evolve.
  • Invest in institutions and infrastructures that shape decision environments: transparency tools, liability rules, certification regimes, data‑sharing frameworks, and capacity building for decentralized oversight.
• Political economy and distributional considerations:
  • Anticipate strategic behavior: actors will reinterpret problem constructions to align with interests; policy must account for contestation and capture by shaping the governance of problem definition (who sets standards, who adjudicates).
  • Consider path dependence: early policy choices shape emergent management norms and market structure; economic analysis should incorporate dynamic and coordination effects.
• Research agenda recommendations:
  • Empirically map the “social management system” around AI (networks of decision makers, incentives, routines).
  • Study how specific policy instruments change firms’ internal decision rules (e.g., governance processes for model deployment) and cross‑actor coordination.
  • Explore mechanisms to make second‑order governance legible and accountable (e.g., audits, meta‑regulatory indicators).

Limitations: because the paper is conceptual, empirical validation is needed to test whether treating policy as second‑order governance improves predictive power or policy outcomes in AI contexts.